Evaluating alternative methods for modeling trap efficiencies of out-migrating juvenile salmonids

Abstract Objective We aimed to compare two machine learning approaches—boosted beta regression (BBR) and beta mixed model forest (BMF)—to a Bayesian mixed-effects beta regression (BME) for the prediction of rotary screw trap (RST) efficiency for out-migrating juvenile salmonids from environmental covariates. Methods We identified two machine learning approaches that shared the ability to model overdispersed probabilities. We compared the BBR and BMF machine learning models to a BME model to evaluate precision in detection probability prediction and model performance on bias in parameter estimation. We tested our three candidate models using a simulation study to understand the specific advantages and disadvantages of each when the data set was increasingly sparse and the capture probabilities were realistically small. We then applied the models to a case study of RST data from the Klamath River in California, United States. Results The BME and BMF outperformed BBR in all simulated scenarios, although the BMF displayed poor explanatory power. In the case study, the BME and BMF identified environmental covariates that predicted RST efficiency. Conclusions Using the BME as a benchmark for comparing machine learning approaches to trap efficiency modeling, our simulations and case study demonstrated that the BMF performed well and is a viable modeling approach with strong predictive power. The BME model would be the preferred modeling approach when its strong explanatory power is desired.